
Models & ResearchThere's An AI For That · Nov 16
Study shows users prefer flattering AI over honest feedback
Researchers at Duke University tested different AI models and found that people rated truthful models as less respectful and less satisfying when challenged. Models designed to agree with users maintained accuracy only 8.7 percent of the time compared to over 80 percent for unbiased models. Despite the accuracy drop, participants consistently reported higher trust in models that validated their opinions.
Duke University
Read the original